Abstract
Using federated learning, which is a distributed machine learning approach, a machine learning model can train on a distributed data set without having to transfer any data between computers. Instead of using a centralised server for training, the model uses data stored locally on the device itself. After that, the server uses this model to create a jointly trained model. Federated learning asserts that privacy is preserved because no data is sent. Botnet attacks are detected using on-device decentralised traffic statistics and a deep autoencoder. This proposed federated learning approach addresses privacy and security concerns about data privacy and security rather than allowing data to be transferred or relocated off the network edge. In order to get the intended results of a previously centralised machine learning technique while also increasing data security, computation will be shifted to the edge layer. Up to 98% accuracy is achieved in anomaly detection with our proposed model using features like MAC IP and source/destination/IP for training. Our solution outperforms a standard centrally managed system in terms of attack detection accuracy, according to our comparative performance analysis.